Atmospheric Pollution Research 13 (2022) 101249

Contents lists available at ScienceDirect

Atmospheric Pollution Research

journal homepage: www.elsevier.com/locate/apr

A sensitivity analysis applied to SPRAY and CALPUFF models when

simulating dispersion from industrial fires
Francesca Tagliaferri, Marzio Invernizzi *, Laura Capelli
Politecnico di Milano, Department of Chemistry, Materials and Chemical Engineering “Giulio Natta” - Piazza Leonardo da Vinci 32, 20133, Milano, Italy

A R T I C L E I N F O A B S T R A C T

Keywords: This paper discusses a hypothetical case study in which SPRAY and CALPUFF dispersion models are applied to
Atmospheric dispersion modelling the simulation of an incidental fire. For this type of accident, source features are typically not directly
Sensitivity analysis measurable, thus making their definition critical. The choice of some model-specific parameters is another
Environmental impact
critical issue, since clear indications are rarely available in guidelines. The aim of this work is to compare how
Models comparison
Source characterization
pollutant concentrations simulated with the two models are affected by changing these two sets of data (i.e.
Fire simulation parameters related to the emission source and model specific parameters), thus performing a sensitivity study to
identify the most influential variables. The most relevant outcome is that sensitivities of the two models are
generally comparable, except for the source diameter: if the SPRAY model is applied with the specific fire source
option, then the concentrations result almost independent from this parameter. Conversely, when considering
other source-types, the concentrations vary up to ± 60% within the selected uncertainty range.

1. Introduction this tendency is expected to increase with the continuous improvement

At the very beginning of the oil industry, accidental fires were a
matter of common occurrence, often entailing disastrous effects at pe­
troleum refining plants. Today, however, because of the development of
specific practices of fire prevention and extinguishing, fires are less
frequent. Nonetheless, safety measures cannot completely prevent this
type of accidents (Shie and Chan, 2013; Sonnemans et al., 2010). In
addition, when they occur, they often have devastating consequences
(Nivolianitou et al., 2006; Zheng and Chen, 2011). First, in terms of
economic losses: a small accident may cause million-dollar property
losses as well as some days of production interruption (Chang and Lin,
2006). The second issue concerns the environmental damages caused by
a fire on air quality, soil and water (Langmann et al., 2009; Weichenthal
et al., 2015). Indeed, the consequences on people’s lives and health is
the matter of greatest concern (Griffiths et al., 2018).
The growing interest in monitoring air quality and assessing health
risks makes the evaluation of the consequences of a fire a key issue.
Atmospheric dispersion models, which simulate the spatial distribution
of pollutants, represent an increasingly widespread tool for this type of
evaluations (Leelőssy et al., 2014). The use of numerical modelling in
the field of industrial fire accidents has become common nowadays, and
of the simulation tools.
Despite the growing number of applications of the modelling
approach to evaluate the consequences of fires (Adame et al., 2018;
Henderson et al., 2008), on a regulatory level, precise guidelines
regarding the type of model and the setting of the model parameters are
not available.
To perform a modelling study, input data relevant to the simulated
domain (mapping, orography and land use) and meteorological data are
required. Furthermore, data concerning the emission scenarios are
needed, i.e. information concerning the emitted species (e.g., emission
factors), the source geometry, and its location in the domain. Finally,
model-specific parameters, which are different depending on the
dispersion model used, must be implemented to perform the
simulations.
Despite the great advantages associated with the use of dispersion
models for atmospheric impact assessment, there are still some impor­
tant issues related to the uncertainty of these models (Chettouh et al.,
2014; Chutia et al., 2014; Russell and Dennis, 2000; Seibert, 2000).
Accuracy of results obtained from mathematical models is often hardly
estimated, because of the presence of uncertainties in the input data.
Indeed, each of the input datasets represents a possible source of error

Peer review under responsibility of Turkish National Committee for Air Pollution Research and Control.
* Corresponding author.
E-mail address: marzio.invernizzi@polimi.it (M. Invernizzi).

https://doi.org/10.1016/j.apr.2021.101249
Received 6 June 2021; Received in revised form 26 October 2021; Accepted 26 October 2021
Available online 3 November 2021
1309-1042/© 2021 Turkish National Committee for Air Pollution Research and Control. Production and hosting by Elsevier B.V. All rights reserved.
F. Tagliaferri et al.
(Holnicki and Nahorski, 2015). In view of this, a sensitivity analysis is
important to explore and quantify the impact of possible changes in
input data on the model outputs.
The aim of this paper is, based on a hypothetical case-study of a
refinery fire, to discuss the sensitivity of two dispersion models applied
to simulate the dispersion of atmospheric pollutants, with the purpose of
investigating the most influential model input data.
As hypothetical case-study, we decided to consider a relatively small
fire, involving a portion of gas oil treatment unit. The reason for this
choice is that the application of dispersion models is consistent in such
conditions, whereas for large catastrophic explosions the lack of repre­
sentation of the explosive phase or plume buoyancy in many models
limits their application.
This work discusses the influence of possible errors in the source data
and in model-specific parameters on the results. The first objective is
particularly interesting because of the high uncertainty in the estimation
of source term parameters for fires (Daly et al., 2012). Indeed, in the case
of fires, source geometrical features are hardly directly measurable, but
they have to be estimated using specific correlations. In our case study,
the source term parameters (i.e. height, diameter and temperature) were
varied within a fairly limited range, since, as previously mentioned, we
decided to simulate a localized fire involving a single equipment.
Furthermore, the choice of some model-specific parameters is often
critical, because clear indications are rarely available in literature or in
the specific model user’s guides, and so their definition is generally left
to the professional judgement of the model user.
The work described in this paper uses two different approaches to
assess the variability in the model results. First, starting from a “base-
case”, defined as the most representative scenario for the case study,
different “plausible” emissive scenarios are investigated. Each of these
scenarios is characterised by a “macroscopic variation” of a single
parameter within the hypothesized uncertainty range. In this study, the
term “macroscopic variation” refers to the extremes of the considered
range of variation of each parameter. The results obtained for each
scenario are compared to the results of the “base-case”. This way, it is
possible to evaluate the effect caused by a wrong estimation of an input
datum, which is crucially important especially in the case of environ­
mental and health impact assessments. This first approach allows to
identify the most influential parameters on the model outputs.
The second approach aims to further analyse the effect of these pa­
rameters by considering “microscopic” variations thereof, i.e. a small
Thus, this second investigation is intended to evaluate effectively the
model sensitivity to the different parameters, by imposing the same
small percent perturbation (compared to the reference base-case) to all,
regardless of the reasonable range of variability.
However, it should be highlighted that this paper does not aim to
present an exhaustive numerical sensitivity study. The adopted
approach highlights the importance of having a broader view of the
issue relevant to the model sensitivity, thereby carrying out additional
investigations focused on those parameters that potentially generate the
highest variations in the results.
It is worth noting that this sensitivity study has not been validated
with experimental field measurements, because it is a hypothetical
event. Conversely, in real conditions, a model validation is strongly
suggested to evaluate the model accuracy and to select properly the
reference base-case as the most representative scenario.
For the purposes of the study, two models have been selected, i.e. the
Lagrangian particle model SPRAY and the puff model CALPUFF (Elbir,
2003; Elbir et al., 2010; Holnicki et al., 2016; Rzeszutek, 2019). Indeed,
Lagrangian particle models and puff models currently represent the most
common tools to simulate pollutants dispersion from fires (Adame et al.,
2018; Ainslie and Jackson, 2009; Henderson et al., 2008). These models
have been chosen because there is a large variety of studies that prove
their validity (Invernizzi et al., 2021), although, as previously
mentioned, there are no specific indications on a regulatory level.
It should be pointed out that the investigation of the most influential
2
Atmospheric Pollution Research 13 (2022) 101249
parameters cannot be generalized to any model or to any situation, but it
refers specifically to the selected case-study. This work does not aim to
suggest how to perform a sensitivity study applicable to any situation,
nor addresses model developers interested in developing or upgrading
software. Actually, it highlights the importance of investigating the
possible range of variation of the input data to identify the most influ­
ential variables, and it may represent a sort of guideline for model users
who have to deal with the implementation of similar case-studies.
Even though, in recent years, the use of dispersion models to simu­
late pollutant dispersion into the atmosphere is continuously increasing,
to the best of our knowledge, in literature there are few studies relevant
to model comparison in terms of sensitivity (Antonioni et al., 2012;
Björnham et al., 2020; Devenish et al., 2012; Gant et al., 2013; Srinivas
et al., 2016). Also, the novelty of this work is related to the approach
followed to evaluate the model sensitivity. Indeed, the paper is not
limited to investigate the influence of the source term parameters on the
model outputs, but it also considers model-specific parameters. The
latter are particularly interesting because for many of them, clear in­
dications on their setting are lacking both in the literature and in the
model user’s guides, thus making it a critical issue for the modelist.
Moreover, differently from other papers, this work proposes a novel dual
approach to address the issues that are relevant to the definition of the
input data by investigating the model sensitivity on two different levels.
2. Materials and methods
2.1. Case – study description
The hypothesized case study regards an incidental fire in an oil re­
finery: the event is supposed to involve a portion of the gas oil treatment
unit.
In our hypothesis, the fire lasts 3 h. In real cases, the duration is a
fundamental point to define the case study and must be evaluated based
on the statements of people who were on the spot.
To optimize the choice of the geographic simulation domain, the
plume direction should be considered. For the selected case study, a
rectangular domain of 25 × 25 km has been identified with a mesh grid
of 250 m. Then, assuming that the plume evolves in south-western di­
rection, the source has been located at the north-eastern corner of the
domain.
In addition, for a more precise analysis, some discrete receptors
should be positioned in places considered of particular interest to esti­
mate the pollutants concentration resulting from the incidental fire (e.g.,
hospitals, schools, city hall). In the identified scenario some receptors
have been located as to be representative of possible places of interest.
2.2. Selection of the dispersion model
A key element for an effective dispersion modelling study is to
choose an appropriate modelling tool to match the scale of impact and
the complexity of the emission scenario. The choice of the most suitable
model can be based on the study of the scientific literature and the
analysis of the technical legislation, which are both useful in order to
understand the features of each model.
The scientific literature is quite deficient in terms of studies con­
cerning accidental fires in oil refineries. Also, even though there are
some studies comparing Lagrangian particle models and puff models
(Invernizzi et al., 2020; Ravina et al., 2020; Souto et al., 2001), to the
best of our knowledge, papers comparing their sensitivity to input data
are lacking.
Nonetheless, there are several reviews (Holmes and Morawska,
2006; Leelőssy et al., 2014; Sinha et al., 2004), which illustrate, in a
rather comprehensive way, the different types of models that can be
used for the simulation of pollutants dispersion. Those reviews typically
describe the model’s features and discuss advantages and disadvantages
according to the kind of application.
F. Tagliaferri et al.
In particular, according to the literature, Gaussian plume models
(Islam, 1999; Liu et al., 2015), Lagrangian puff models (Jung et al.,
2003), Lagrangian particle models (Cécé et al., 2016; Santiago and
Martín, 2008), Eulerian grid models (Kota et al., 2013; Seland and
Iversen, 1999) and fluid dynamics models (Leelőssy et al., 2014; Mar­
katos et al., 2009) may be used to simulate pollutants dispersion. A
detailed description of the different classes of dispersion models avail­
able is out of the scope of this paper.
Here, some considerations have been made in order to choose the
most suitable model for the selected case study:
• The greatest advantage of Gaussian models is that they have an
extremely fast, almost immediate, response time. However, the use
of very simple models, such as Gaussians, is not advisable in case of
large simulation domains (Daly et al., 2012), because they consider
steady state conditions. Thus, they cannot adequately describe the
dispersive phenomenon, since the meteorological condition of one
point of the domain is not representative of the wind field variations
over the entire domain.
• Eulerian and fluid dynamics models are very advanced simulators,
but at the same time they are very complex and require a long
computational time.
• Fluid dynamics models fare suggested, or even necessary, when the
dispersion occurs in urban areas, where the influence of buildings on
the dispersion is dominant, or where the scale to consider is the so-
called meteorological microscale (<1 km) (Bhuiyan and Naser,
2015; Jang et al., 2015; Mishra and Wehrstedt, 2015; Novozhilov,
2001). However, the considered case study has a simulation domain
larger than 1 km and the simulated site is not located in an urban
area, making such complex models unneeded.
For these reasons, a Lagrangian puff model (i.e. CALPUFF) and a
Lagrangian particle model (i.e. SPRAY) have been chosen: they repre­
sent a compromise between reasonable accuracy and manageable
computational time.
Furthermore, the use of these models for the selected case study is
compliant with the Italian technical standards on the matter (UNI
10796, 2009; UNI 10964, 2009). These standards define the scenarios
for the implementation of different models, suggesting the best model
for each situation.
Finally, the available scientific literature also supports this choice:
past studies prove the suitability of these models in similar cases. For
instance, there are some examples of studies carried out using puff
models, and specifically CALPUFF, for the simulation of pollutant
dispersion from fires (Ainslie and Jackson, 2009; Henderson et al.,
2008).
On the other hand, there are fewer articles regarding the application
of SPRAY to the simulation of fires, presumably because it is a more
recently developed model. However, the bibliographic research proves
the suitability of the Lagrangian models to simulate fires: the model
behaviour is usually compared with measured data to test its goodness,
showing in most cases a good accuracy (Adame et al., 2018; Liu et al.,
2018).
2.3. Meteorological and orographical features of the model
The meteorological data used for the simulations are three-
dimensional prognostic WRF data purchased from Lakes Environ­
mental. Each model processes the WRF data using the model-specific
meteorological tools (i.e. SWIFT for SPRAY and CALMET for CAL­
PUFF), which are diagnostic “mass consistent” models. They generate
3D wind fields inside the meteorological domain, which has been set
equal to the computational grid. For the simulations, the meteorological
conditions were chosen to be representative of a neutral winter morning
with cloud cover.
The model also requires the site orography as input data. In this case,
3
Atmospheric Pollution Research 13 (2022) 101249
the simulated area is characterized by a flat land.
2.4. Definition of base-case
Before starting a modelling study, it is necessary to quantify some
parameters needed as input data.
In the case of a “real fire”, the first step is the quantification of the
amount of fuel burnt. This can be estimated by a mass balance around
the involved equipment.
Then, for an incidental fire, the involved compounds and their
emission factors shall be defined. The scientific and technical literature
states that the most common pollutants associated with gas oil com­
bustion are CO2, CO, generic unburnt hydrocarbons (CH), particulate
matter (PM), SOX and NOX (Booher and Janke, 1997; Lemieux et al.,
2004).
The SFPE Handbook of Fire Protection Engineering (DiNenno et al.,
2002) has been chosen for the estimation of emission factors, because it
has been considered to be an authoritative and reliable reference. More
in detail, Tables 3–4.14 “Yields of Fire Products and Chemical,
Convective, and Radiative Heats of Combustion for Well-Ventilated
Fires” has been considered for the definition of the emission factors of
the CO2, CO, unburnt hydrocarbons (CH) and particulate matter (PM)
emitted during the combustion of different fuels. Among all the listed
species, the generic “Hydrocarbon” and the kerosene have been
considered, thus obtaining the following emission factors (ton/ton): 2.7
for CO2, 0.02 for CO, 0.007 for CH, and 0.05 for PM.
The SFPE Handbook of Fire Protection Engineering does not report
emission factors for NOX emissions. Therefore, the emission factor for
NOx, which was set equal to 0.01 (ton/ton), was defined using the AP-42
database of the US EPA, 1998.
For SOX we assumed the complete (stoichiometric) conversion to SO2
of the elemental sulphur (S) in the original fuel, which was hypothesized
to be 100 ppm, thus giving an emission factor of 0.002 (ton/ton).
The emission rate for each compound can be calculated as the
product of its emission factor and the amount of fuel burnt, divided by
the event duration.
Furthermore, the model requires the geometrical features of the
simulated source. Therefore, diameter and height have to be evaluated
from surveys and technical documentation.
In addition to the geometrical characterization of the source, some
physical parameters have to be defined. The fire smoke rise velocity is
derived from the Ingason correlation (Ingason and Li, 2015):
⎛ ⎞13
/
⎜
w=⎝
g⋅q ⎟
⎠ (1)
D 2⋅ρ0 ⋅CP ⋅Ta
/
where Ta is the ambient temperature, ρ0 the air density, cp the specific
heat of air at constant pressure, g is the acceleration of gravity, q the heat
release and D the source diameter.
The Ingason correlation requires the estimation of the heat release,
which may be derived from the Babrauskas correlation (DiNenno et al.,
2002):
( )
q = m′′∞ ⋅ΔhC,eff 1 − exp− (kβD) A (2)
where Δhc,eff is the net heat of combustion, A is the source area, D is the
diameter, and kβ and m′′∞ are empirical constants available in literature
for a number of common fuels. The values for the fuel of interest are
obtained from DiNenno et al. (2002), Tables 3–1.13 “Pool Burning:
Thermochemical and Empirical Constants for a Number of Common
Organic Fuels”, which reports the empirical constants for the most
common organic fuels. Since specific data for gas oil are not available,
the fuel that has been considered most representative among those listed
is JP-5 (Δhc,eff = 43 MJ/kg; kβ = 1.6 m− 1; m′′∞ = 0.054 kg/(m2s)).
Finally, it is necessary to define the fire temperature. To do this, the
F. Tagliaferri et al.
hydrocarbon fire curve, reported in BS EN 1363-2, 1999 showing the
trend of the temperature as a function of time, has been used.
The maximum achievable temperature of about 1100 ◦ C, has been
considered for the hypothesized case study since it is rapidly achieved
after a few minutes.
Based on these evaluations, a reference “base-case” was defined by
setting the input parameters considered as most representative of the
hypothesized emission scenario.
Its characteristic parameters are shown in Table 1.
2.5. Source types and plume rise computation
One of the investigated parameters is the modelled source type. For
both CALPUFF and SPRAY, the fire is modelled by applying the specific
source type suggested for fires (i.e. the buoyant area source for CAL­
PUFF, and the fire for SPRAY) and then compared with the model results
obtained by assimilating the fire to a point source (i.e. stack).
According to the CALPUFF User’s Guide (Scire et al., 2000), the
buoyant area source is the most appropriate tool to simulate fires.
Therefore, for the CALPUFF simulations, the point source has been
considered only for the “base-case”, because of its unsuitability to model
the selected event. Despite the indications of the User’s Guide, the point
source has been implemented in the CALPUFF “base-case” to highlight
the differences in the ground concentrations in comparison to the
buoyant area source, as will be discussed in Paragraph 3.1, and,
consequently, to underline the importance of using the specific fire
options available in the models.
For each type of source, the dispersion model simulates the plume
rise mechanism according to a different scheme. When considering the
fire as a point source (without any fire-specific option), both CALPUFF
and SPRAY calculate the buoyancy flux according to the Briggs equation
(Tinarelli, 2017), which is proportional to the square of the source
radius:
T − Ta
Fb = gr2 w0 (3)
Ta
where g is the acceleration of gravity, r the source radius, w0 the effluent
exit velocity, Ta the ambient temperature and T the exit smoke tem­
perature.
When using the CALPUFF - buoyant area source model, the radiative
heat loss from the plume to the ambient air can be estimated using the
following equation (Scire et al., 2000):
/
q 2 ( )
r = − 2εσ r T 4 − Ta4 cp (4)
cp
where q is the radiative heat loss, cp the specific heat of the ambient air, r
the source radius, σ the Stefan-Boltzmann constant, ε the emissivity, T
the plume temperature and Ta the ambient temperature. Here, an in­
crease of the radius implies a reduction of the heat losses. Consequently,
the plume rise increases and the pollutant concentrations decrease.
On the other hand, if the SPRAY model is used in combination with
the specific fire option, the equation used for the buoyancy calculation,
in which neither velocity nor radius appear, results in a buoyancy flux
not affected by the source diameter (Tinarelli, 2017):
gP
Fb = ε (5)
πcp ρair Ta
Table 1
Source term parameters for the base-case. D = source diameter; T = plume temperature; H = source height; v = exit velocity; q = heat release; Quantity = amount of
fuel burnt.
Scenario D T H v Q Quantity
[m] [K] [m] [m/s] [kW]
BASE 5 1373 15 8.16 46,354
Atmospheric Pollution Research 13 (2022) 101249
where ε = 0.7 represents the reduction term due to radiation, P is the
energy/time, g the acceleration of gravity, cp and ρair the specific heat
and the density of the ambient air, and Ta is the ambient temperature.
2.6. Definition of alternative cases for the evaluation of the effects of
source characterization
Starting from the “base-case”, it has been decided to investigate
alternative emission scenarios by changing the most critical geometrical
parameters of the source within a reasonable range of variation. Indeed,
the estimation of these variables is characterized by high uncertainty
because of the impossibility of measuring them directly during the
event.
In addition to the source geometrical parameters, other variables
associated with the definition of the emission scenario have been
investigated (e.g., temperature and amount of fuel burnt).
Under real conditions of a refinery fire, in order to define the alter­
native scenarios, it is advisable to first validate the base-case, with the
purpose to verify that this is effectively the most representative and
realistic scenario. However, it shall be considered that this is usually
very complex in case of accidental fires.
The alternative scenarios defined for the study are shown in Table 2,
with the numbers in bold representing the variables changed in the
alternative scenarios.
2.7. Evaluation of the effects of model-specific parameters
2.7.1. CALPUFF
As far as CALPUFF is concerned, the only model-specific parameter
that has been investigated is the initial vertical dispersion coefficient σz0,
which defines the initial dimension of the puff in the vertical direction.
Concerning the base-case scenario, as suggested in the user’s guide for
Gaussian plume models (US EPA, 1995), σz0 is evaluated as follows:
σ z0 = H/2.15 (6)
where H is the source height.
This correlation is recommended for surface-based sources, as it is
the case in the modelled case-study.
However, in literature, other correlations are reported, which can be
adopted depending on the source elevation or the presence of adjacent
buildings. More in detail, in case of elevated source located on or
adjacent to a building, it is suggested to divide the building height by a
factor of 2.15. In case of an elevated source not adjacent to a building, it
is recommended to divide the vertical dimension of the source by 4.3
(US EPA, 1995). Accordingly, to run the alternative scenario, σz0 is set
equal to the source height divided by 4.3.
2.7.2. SPRAY
SPRAY is a more advanced software: it requires the definition of
several parameters, whose estimation is not trivial, due to the absence of
specific indications.
2.7.2.1. Height of the first layer. Differently from CALPUFF, which
provides the output concentration in a gridded surface at the ground, the
Lagrangian particle model needs the height of the first layer as input
datum: it is the height of the first cell above ground used by the model to
PM CO CO2 NOx SO2 HC
[ton] [g/s] [g/s] [g/s] [g/s] [g/s] [g/s]
11.2 51.8 20.7 2800 10.4 0.2 7.3
4
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249

Table 2
Alternative scenarios for the source term parameters.
Scenario D T H v Quantity PM CO CO2 NOx SO2 HC

[m] [K] [m] [m/s] [ton] [g/s] [g/s] [g/s] [g/s] [g/s] [g/s]

A1 10 1373 15 6.21 11.2 51.852 20.741 2800 10.37 0.207 7.259

A2 3.5 1373 15 9.17 11.2 51.852 20.741 2800 10.37 0.207 7.259
H1 5 1373 20 8.16 11.2 51.852 20.741 2800 10.37 0.207 7.259
T1 5 1273 15 8.16 11.2 51.852 20.741 2800 10.37 0.207 7.259
T2 5 1473 15 8.16 11.2 51.852 20.741 2800 10.37 0.207 7.259
Q2 5 1373 15 10.28 22.4 103.703 41.481 5600 20.74 0.4148 14.518
Q2A1 10 1373 15 7.86 22.4 103.703 41.481 5600 20.74 0.4148 14.518
Q5A1 10 1373 15 10.68 56.0 259.260 103.71 14,000 51.85 1.035 36.295

It is worth underlining that, when using the SPRAY - fire model, scenarios T1 and T2 have not been considered, since the fire temperature is not an input parameter
required by the software.

compute the concentration in any point P (x; y; z; t) at time t. To do this,
the model computes the pollutant concentration considering a “sam­
pling volume” having the grid step dimensions in x and y direction, and
the third dimension z is the height of the first layer.
The value attributed to this parameter for the base-case is 10 m: on
one hand, in order to evaluate the ground level concentration, the first
layer height should be sufficiently low and, on the other hand, a too low
value is not advisable in order to limit the influence of the mechanical
turbulence, the effect of which is more pronounced in the vicinity of the
terrain, progressively decreasing moving far from the ground.
In addition, the surface roughness shall be considered for the eval­
uation of this parameter. The roughness length z0 represents the height
where the wind speed becomes zero (no-slip condition) and it is related
to the terrain features: depending on the land use type, different values
for z0 are suggested. In particular, in case of urban areas, SPRAY
(ARIANET, 2011) sets this parameter to 1 m. Hanna and Britter (2002)
suggested that the ratio between z0 and the obstacle height Hr can be
estimated according to a simple rule of thumb:
z0
= 0.1
Hr
It follows that the parameter Hr assumes a value of 10 m. Also, Hanna
and Britter (2002) highlighted that for a typical urban or industrial site
an average building height of 10 m is a reasonable estimation. Since one
of the main purposes of an atmospheric dispersion model is the assess­
ment of the impact on people, it is reasonable to consider the height up
to which the concentration estimation is of interest (i.e. the height of the
first layer) of 10 m.
2.7.2.2. Δz. Another model-specific parameter that has to be defined is
Δz, i.e. the vertical dimension of the “emission parallelepiped”. SPRAY
generates particles uniformly distributed on a “terrain following”
parallelepiped centred in P (X0, Y0, Z0), which are the coordinates of the
emission region centre of gravity, the vertical dimension of which is Δz
(Tinarelli, 2017).
In other words, this parallelepiped can be thought of as a box in
which the particles initially appear. Thus, they are released in a vertical
region ranging from Z0– Δz/2 and Z0 + Δz/2, with Z0 coincident with
area (building downwash) resulting in high concentrations immediately
downwind of the source. Therefore, to reproduce the emission region, a
conservative value equal to twice the source height (30 m) has been
defined.
2.7.2.3. Number of particles. Another investigated parameter is related
to the stochastic description of the Lagrangian model. The air is
described as a set of parcels that move according to two different
mechanisms: advection and turbulent motion. The irregular and highly
variable nature of the main parameters describing the motion of a
molecule in the air makes it not possible to use exact values of them in
any practical problem, since a fully deterministic approach is almost
impossible.
The SPRAY model requires the definition of the number of particles
used for the simulations; whereby each particle represents a discrete
amount of pollutant. For this variable, a suitable value has been selected
in retrospect, after running some simulations with different numbers of
particles. To identify a reasonable value, a compromise between good
accuracy in the results and manageable computational time has been
(7) considered, leading to a choice of about 3 million of particles emitted in
the 3 h of simulation. To ensure a detailed description of the particles
motion, a sufficiently high number of particles is required. Indeed, the
smaller the sample size is, the more outliers may skew the findings. In
other words, the particles represent the air parcels. Thus, each particle
has a random contribution of motion that has to be described consid­
ering a stochastic approach. If the particles number is high enough, an
average behaviour can be identified, minimizing the discrete contribu­
tion of each particle. Conversely, if few particles are considered, there is
the risk that the outliers are heavier. In other words, the number of
particles should be high enough as to ensure that the results do not show
any statistically meaningful difference when changing this variable.
Table 3 shows the model-specific parameters for SPRAY and
Table 3
SPRAY and CALPUFF model specific parameters for base-case and alternative
scenarios.
SPRAY CALPUFF

the source height. In the Supplementary Material S1 a sketch of the Scenario Δz H1 Layer
◦ Particle number Scenario σZ0
emission parallelepiped is shown, highlighting its vertical dimension Δz, [m] [m] [m]
and its position with respect to the emission source. BASE 30 10 2′ 989′ 440 BASE 6.98 ( = H/2.15)
This variable describes the initial condition of the emission and it Δz2 15 10 2′ 989′ 440 σZ0,1 3.49 ( = H/4.3)
should be defined as to reproduce the geometrical features of the Δz3 20 10 2′ 989′ 440
25 10 2′ 989′ 440
emission region. The “geometrical features of the emission region” refers
Δz4
H_1L 30 4 2′ 989′ 440
not only to the effective source dimensions, but also to the dynamic PARTICLE1 30 10 151′200
effects affecting the emission. For instance, because of the configuration PARTICLE2 30 10 749′520
of the stack or of the adjacent buildings, the plume may not rise freely in PARTICLE3 30 10 2′492′640
PARTICLE4 30 10 2′989′440
the atmosphere: some aerodynamic effects due to the way the wind
PARTICLE5 30 10 7′471′440
moves around adjacent buildings and the stack can force the plume to­ PARTICLE6 30 10 10′670′400
wards the ground instead of allowing it to rise. Depending on the stack PARTICLE7 30 10 14′938′560
height, it may be possible for the plume to be pulled down into this wake PARTICLE8 30 10 24′896′160

5
F. Tagliaferri et al.
CALPUFF selected for the base-case and the alternative scenarios, with
the numbers in bold representing the variables changed in the alterna­
tive scenarios.
In order to investigate the effect of the particles number, the case
with the highest number of particles (i.e. 74′ 682′ 000 particles emitted in
3 h, from the SPRAY default release option of 34’575 particles in 5 s) has
been considered as a reference, whereas the “base-case” with 2′ 989′ 440
particles is treated as an alternative scenario. This is the reason why, in
Table 3, the scenario with 74′ 682′ 000 is not reported, because it is
considered as the reference scenario to compute the sensitivity of the
results with respect to the number of particles in the alternative sce­
narios (PARTICLE1 - PARTICLE8).
2.8. Sensitivity analysis
Once the most influential parameters have been identified by means
of the simulations relevant to the alternative scenarios, a sensitivity
analysis is needed in order to quantify the effects of a “microscopic
variation” of the input variables on the model outputs.
2.8.1. Choice of the approach
The choice of the method to be used for the sensitivity analysis was
based on a deep literature search. As a result, a paper proposing a
sensitivity analysis based on the Taylor series approach by Yegnan et al.
(2002) has been selected, because it well-suits the hypothesized
case-study.
Indeed, Yegnan et al. (2002) adopted this approach to calculate the
sensitivity of ground level concentrations resulting from a short-term
simulation (1 h), as in the case ofour hypothesis. Also, after the defini­
tion of a “base-case” for the modelling scenario, each of the seven input
parameters (many of which are the same investigated in the alternative
scenarios reported in Table 2, e.g., stack height, stack diameter, tem­
perature) has been modified to determine the ones that mostly affect the
output concentrations, i.e. the parameters to which the model is most
sensitive.
In the paper by Yegnan et al. (2002), the “base-case” identified is the
most representative scenario the average values are derived from. Then,
the most sensitive input parameters, i.e. wind speed and temperature,
are perturbed by 1% in both directions with respect to the average
values, and the corresponding change in the output is evaluated. The
sensitivity of the output f’(x) is then computed as:
′
f (x) =
f (x2 ) − f (x1 )
(8)
x2 − x1
where x is the value of the input parameter adopted in the reference
base-case, x1 and x2 are the perturbed input values on either side of x ,
and f(x1 ) and f(x2 ) are the corresponding output values.
Also, to normalize the results by removing the effects of units, the
dimensionless sensitivity index (Gonsamo, 2011; Rodrigues et al., 2013)
is introduced:
x explained in section 2.5. The simulation of PM dispersion has been run
(9)
′ ′
f (x)N = f (x)⋅
y as for an inert gas (no deposition considered). PM was selected as the
where the apex N refers to the normalization of the sensitivity index, and
y is the value of the output resulting from the reference base-case (with x
as input variable).
Rodrigues et al. (2013) defined the normalized sensitivity index of a
variable with respect to a parameter as “the ratio of the relative change
in the variable to the relative change in the parameter”. Thus, it is ob­
tained by multiplying f (x) by the ratio of the parameter value to the
′
model result for the base-case scenario.
2.8.2. Application of the method to the case study
To perform the sensitivity study discussed in the previous paragraph,
the investigated parameters have been perturbed by 1%, 2% and 3% in
6
Atmospheric Pollution Research 13 (2022) 101249
both directions with respect to the value of the base-case, while all the
other parameters were kept unchanged.
Then, the concentration values resulting from these perturbations
have to be computed on a set of discrete receptors. In this case, 40 re­
ceptors placed along the plume axis starting from the source and spaced
250 m from each other were considered.
This sensitivity study has a different purpose compared to the
approach based on the evaluation of the alternative scenarios. In this
case, the perturbations imposed to the investigated variables are not the
extremes of a reasonable range of variability for the selected parameter,
but a “microscopic” perturbation is considered. As a result, the sensi­
tivity analysis is intended to evaluate the numerical model sensitivity to
the investigated variables, regardless of the range of variability thereof.
In particular, this approach allows to:
• Compare effectively the model sensitivity to different parameters, by
imposing the same perturbations to all of them;
• Investigate the model behaviour caused by a perturbation of the
selected parameter, identifying, for instance, the relationship exist­
ing between the input and the output;
• Identify the way the model sensitivity to the selected parameter
changes if moving far from the source;
• Compare SPRAY and CALPUFF sensitivity to the selected parameters,
highlighting which of them is more sensitive to perturbations of the
investigated input data.
One important preliminary consideration concerns the applicability
of this approach to the SPRAY model, whereby the number of particles
may affect the significance of the calculated sensitivity indexes. Indeed,
to obtain reliable results from this test, since the variation applied to the
parameters is significantly low (i.e. 1%–3%), it is important to reduce as
much as possible the influence related to the choice of the number of
particles.
For this reason, it is necessary to adopt a significantly high number of
particles to run these simulations. For the selected case-study, Supple­
mentary Material S2 shows that simulations with 106′ 688′ 880 particles
have negligible variability due to the number of particles. For this
reason, the sensitivity analysis for the SPRAY model is carried out
considering the emission of 106′ 688′ 880 particles.
3. Results and critical discussion
3.1. Results of “base-case” for different source types
The simulation results can be processed with the post processing
tools of each of the models under investigation (i.e. CALPOST for CAL­
PUFF and POSTBIN for SPRAY). This way, ground level concentration
maps representing the pollutants dispersion within the simulation
domain are obtained. As an example of the simulation results, Fig. 1
shows the maximum 1-h concentration maps for the PM resulting from
the base-case simulation in function of the different source types
“tracer” to be treated as reference species for further evaluations, since it
is, among the compounds considered, the one with the highest emission
rate after CO2. Therefore, from this point forward, PM will be referred to
as “Tracer”. It should be highlighted that the reference compound cho­
sen doesn’t affect the forthcoming considerations about model sensi­
tivity or the comparison between different source types: the same
considerations could be applied to any other species considered.
Depending on the source type and the dispersion model used, slightly
different plume directions can be observed. The maps in Fig. 1 differ in
concentration and plume shapes: this can be explained considering the
different wind fields resulting from the application of the different met
pre-processors. Indeed, despite starting from the same raw meteoro­
logical input data, the different meteorological pre-processors (i.e.
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249

Fig. 1. Maximum ground level concentration maps of the tracer species resulting from CALPUFF-buoyant area (a.), SPRAY- fire (b.), CALPUFF-point (c.) and SPRAY-
point (d.). Figure (a.) shows the position of the six discrete receptors discussed below.

CALMET for CALPUFF and SWIFT for SPRAY) elaborate the results in a
slightly different way. As an example, the picture of the wind field
computed by CALMET and SWIFT for the same hour is reported in the
Supplementary Material S3.
The maximum Tracer concentrations calculated by the models on a
set of selected discrete receptors (indicated in Fig. 1(a.)) are reported in
Table 4. More in detail, receptor 1 corresponds to the gridded receptor
where the maximum concentration has been calculated by the two
models inside the simulation domain, receptors (2–5) are representative
Table 4
Maximum Tracer concentration values at selected receptors calculated by
CALPUFF (left) and SPRAY (right).
ID Description CALPUFF SPRAY
of places of public interest and receptor 6 is the point of maximum
concentration at the plant fence line.
Fig. 2 shows the trend of the of maximum 1-h concentration of the
Tracer as a function of the distance from the source. To the purpose, 40
receptors placed along the plume axis starting from the source and
spaced 250 m from each other are considered.
The different trends shown in Fig. 2 can be explained by considering
the different plume rise computations for point sources and for fires/
buoyant area sources. According to the SPRAY model for fires, which
considers a non-complete combustion, there is a cold fraction of parti­
cles that remains unburnt and immediately falls to the ground, without
being dragged into the plume rise. This gives the highest Tracer

Conc. (point Conc. Conc. (point Conc.

source) (buoyant source) (fire)
area)

[μg/m3] [μg/m3] [μg/m3] [μg/

m3]

R_1 MAX DOMAIN 5.12 114.48 12.97 213.2

R_2 RANK 1 0.77 1.05 4.19 4.99
SENSITIVE
R_3 RANK 2 0.23 14.69 4.28 23.6
SENSITIVE
R_4 RANK 3 2.44 11.46 5.67 19.45
SENSITIVE
R_5 RANK 4 3.78 18.83 5.60 22.87
SENSITIVE Fig. 2. Maximum Tracer concentration trend in function of the source distance
R_6 MAX 0.31 19.29 6.64 23.31
for the different combinations of dispersion models and source
FENCELINE
types considered.

7
F. Tagliaferri et al.
concentrations close to the source. As far as CALPUFF is concerned, the
buoyant area source model considers radiative heat losses due to the
high plume temperature near the burning source. Consequently, the
heat flux along the plume trajectory will be reduced, leading to a lower
buoyancy flux. On the contrary, for point sources, the maximized plume
rise leads to very low concentration values close to the emission point,
whereas a concentration peak is observed at the distance where the
plume reaches the ground.
At high distance from the source (>5000m) the maximum Tracer
concentrations computed by the different models tend to become very
similar, giving concentrations ranging from 5 to 9 μg m− 3 at 5000 m
from the source and from 3 to 6 μg m− 3 at 10,000 m from the source.
3.2. Alternative scenarios for source geometrical features and emission
scenario characteristics
Fig. 3 shows, for the different source types, the variability (%) of the
maximum Tracer concentration values resulting at the selected re­
ceptors 1–6 (which are the same as for the base-case, which are reported
in Table 4) from the simulations relevant to the alternative scenarios
were compared to those obtained for the base-case. It is worth recalling
that, for reasons discussed in Section 2.5, the point source simulated by
CALPUFF will not be considered from this moment on.
This investigation shows that the diameter of the source is one of the
most interesting source term parameters, because of its different influ­
ence on the model outputs depending on the considered source type: it
significantly affects the model outputs when using the CALPUFF -
buoyant area source model or the SPRAY - point source model, but it
leads to very low variations when applying the SPRAY model in com­
bination with the specific fire option. Thus, to examine the source
diameter influence more deeply, Table 5 reports the relative variation
(%) relevant to the scenarios A1 and A2, where only the geometric
dimension of the source has been modified if compared to the base case.
Considering CALPUFF and the point source simulated by SPRAY, the
simulations conducted at the boundaries of the uncertainty range for the
source diameter result in significant variations in the maximum
modelled Tracer concentrations at selected receptors. A decrease of the
diameter from 5 m to 3.5 m (scenario A2) generally results in an increase
of the simulated maximum Tracer concentrations of about 50%, whereas
an opposite effect is obtained by increasing the diameter from 5 m to 10
m (scenario A1), generally giving decreased concentrations of about
60%. On the other hand, when using the SPRAY – fire model, the source
diameter does not represent a highly influential variable, giving a
maximum variability of 10% at the selected receptors.
When using the SPRAY - point source model, the effect of the source
can be explained by considering the buoyancy flux computation, per­
formed according to the Briggs equation (see Section 2.5).
The source diameter affects the buoyancy of the plume indirectly
through the radius and the exit velocity, which are required to calculate
the buoyancy flux. More in detail, the diameter has an opposite effect on
these two parameters (see Section 2.5). An increase in the diameter
means a small decrease in the velocity whereas the source radius, which
is squared in the Briggs equation, increases significantly. Therefore, the
dominant term is the second one, leading to an increased buoyancy flux
and a reduction in the ground level concentrations.
When using the CALPUFF - buoyant area source model, the compu­
tation of the radiative heat loss from the plume to the ambient air de­
pends on the source radius. Here, an increase of the radius implies a
reduction of the heat losses. Consequently, the plume rise increases and
the pollutant concentrations decrease.
On the other hand, if the SPRAY model is used in combination with
the specific fire option, the influence of the diameter of the source on the
model outputs turns out to be negligible. From a mathematical point of
view, this can be explained by the equation used (see Section 2.5) for the
calculation of buoyancy flux, where neither velocity nor radius appear,
giving that the buoyancy calculation is not affected by the source
8
Atmospheric Pollution Research 13 (2022) 101249
diameter.
The second investigated parameter is the source height, the influence
of which on the model output, (Fig. 3), is less relevant: for all the
investigated sources, the variability has the same order of magnitude
and it does not significantly affect the model results. Also, depending on
the position of the selected receptor, it may lead to an increase or a
decrease in the maximum Tracer concentration, but in all cases the
variabilities (%) are significantly lower than those produced by the
variation of the source diameter. The same consideration applies to the
model sensitivity to temperature. Its influence on the concentration
values is even lower, giving a maximum variation on the selected re­
ceptors of 5% (Fig. 3).
In the Q2 scenario (Fig. 3), where the amount of fuel burnt is doubled
compared to the reference scenario, the resulting concentration at the
receptors is not doubled, as the calculated variability is always lower
than 100%. The explanation for this behaviour is that, at constant
diameter, by increasing the fuel burnt, the heat released by the fire in­
creases, leading to a rise of the velocity (according to Equation (1)) that
promotes pollutant dispersion.
For the alternative scenarios in which the source diameter and the
amount of fuel burnt have both been modified (Q2A1 and Q5A1), the
variabilities (%) generated by the SPRAY - fire model are significantly
higher than those obtained with the other source types (Fig. 3). Indeed,
when using the fire model, since the diameter is almost irrelevant, the
variation in the Tracer concentration is due only to the increase in the
amount of fuel burnt.
Therefore, from this first evaluation, the most relevant parameters
appear to be the source diameter and the amount of fuel burnt,
although the latter is a parameter that can be usually quantified with
a certain degree of reliability.
3.3. Alternative scenarios for model-specific parameters
3.3.1. CALPUFF: the influence of σ Z0
For the CALPUFF model, the investigated parameter is the initial
vertical dispersion coefficient. In the σZ0,1 scenario this variable has
been halved compared to the base-case. The variability (%) of the
maximum Tracer concentration values resulting at the selected re­
ceptors from the simulation relevant to the σZ0,1 scenario compared to
the base-case is graphically shown in the Supplementary Material S4 and
briefly discussed in this section of the paper.
From the alternative scenario, it turns out that the initial vertical
dispersion coefficient does not significantly affect the ground level
concentrations: when passing from a σZ0 of 6.98 m to a σZ0 of 3.49, the
model results (i.e. the concentrations at the considered receptors) are
subjected to a maximum variation of 9.5% corresponding to the point of
maximum concentration, whereas for the other receptors this variability
is below 2%.
3.3.2. SPRAY: the influence of model-specific parameters
As mentioned in section 2.8, since SPRAY is a more advanced soft­
ware, which requires the definition of several parameters, different
scenarios have been developed to investigate the effect of those vari­
ables whose estimation is not trivial.
The variability (%) of the maximum Tracer concentration values at
the selected receptors (1–6) resulting from the alternative scenarios
relevant to the SPRAY specific parameters (except for the number of
particles number, which will be discussed later) compared to the base-
case has been investigated. In the Supplementary Material S4, a figure
showing the % variability obtained at discrete receptors for the alter­
native scenarios is reported.
A first consideration concerns the percent variations resulting from
the modifications of the model-specific parameters compared to those of
the source term parameters. None of the variations of the model-specific
parameters leads to significant alteration of the model outputs, as it is
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249

Fig. 3. % variation of the maximum PM concentration values at the selected receptors resulting from the simulations of the alternative emission scenarios compared
to the reference base-case for the different source types.

9
F. Tagliaferri et al. Atmospheric Pollution Research 13 (2022) 101249

Table 5 observed for the change of the vertical dimension of the “emission
% variation of the maximum Tracer concentration values at the selected re­ parallelepiped”: for the scenarios Δz2, Δz3 and Δz4, the SPRAY - point
ceptors resulting from the A1 and A2 scenarios compared to the reference base- source model always shows a percent variation higher than the fire
case for the different combinations of models and source types considered. model. Thus, this behaviour does not seem to be attributable to the
ID CALPUFF (Buoyant Area) SPRAY (Point) SPRAY (Fire) position of the selected receptors but it is a general feature resulting
A1 A2 A1 A2 A1 A2 from the different way to model the source. Indeed, a change in Δz,
which represents the vertical dimension of a “box” centred in the
1 − 48.10% 42.10% 52.40% 47.60% 10.20% − 0.80%
emission region centre of gravity in which the particles initially appear
−
2 − 61.90% 7.60% − 24.80% 8.10% 2.00% − 2.40%
3 − 40.50% 35.70% − 64.50% 50.90% 2.50% 0.20% (see Section 2.7.2.2), means a change in the dimension of the region
4 32.90% − 53.30% − 39.00% 24.70% 2.10% 1.00% from which the particles start to rise up due to the buoyancy. Thus, the
5 − 67.10% 74.40% − 38.60% 23.20% 1.10% 0.30% plume rise is affected by this variable in the sense that the “idealized”
6 − 45.70% 21.00% 59.90% 49.90% 1.30% 0.30%
−
plume containing the particles has a different initial shape and dimen­
sion according to this parameter. On the other hand, the fire model
the case for the variations applied to the source diameter or the amount considers a portion of emitted particles with no buoyancy flux: the rapid
of fuel. downfall of these particles makes this kind of simulated source less
Indeed, both the SPRAY – point source model and the SPRAY - fire dependent on the variable Δz.
model show, in correspondence of almost all the receptors considered Another investigated parameter is the number of particles, which
and for almost all the investigated parameters, a variability lower than determines the variabilities (%) reported in Fig. 4. It is worth recalling
10%. In addition, comparing the two source types, their response to the that, when analysing the output variations due to the particles number,
input variation is very similar. The most different behaviour is the one the variabilities are not referred to the base-case, but to a different
reference simulation involving the highest amount of particles

Fig. 4. % variation of the maximum Tracer concentration values at the selected receptors resulting from the simulations of the alternative cases for the particles
number (for SPRAY – point and SPRAY – fire) compared to the reference base-case.

10
F. Tagliaferri et al.
considered (i.e. 74′ 682′ 000 particles emitted in 3 h). The higher the
number of particles, the lower the stochastic variability of the results.
As expected, for both source types, an insufficient number of parti­
cles leads to discrepancies in the results, whereas the percent variation
produced by the number of particles tends to decrease when the input
parameter becomes closer to the maximum number of particles
considered. In particular, for scenario PARTIC8, involving the emission
of 24′ 896′ 160 particles, the percent variability is approximately 1% in
all the selected receptors. Considering scenario PARTIC4, which implies
the release of 2′ 989′ 440 particles, a maximum variability of 3% has been
found. Considering the purpose of the first part of this paper, which aims
to investigate the model response to “macroscopic variations” in the
input data, a “distorted” result of 3% is considered acceptable.
3.4. Sensitivity analysis
The development of alternative scenarios enables to identify the
most influential variables. Looking at the relative variations (%) ana­
lysed for each scenario in the previous paragraphs, the most critical
input datum turns out to be the source diameter (except for the SPRAY -
fire model). Although the amount of fuel burnt also significantly affects
the ground level concentrations, in the case of real fires, this is usually a
parameter that can be quantified with a certain degree of reliability.
That is why the sensitivity analysis, described in Section 2.8, has been
applied only to the source diameter.
More in detail, the analysis has been applied to the point source
simulated by SPRAY and the buoyant area source of CALPUFF, where it
produces the main effects (see Section 3.2).
As already mentioned in Section 2.8.2, the source diameter has been
perturbed by 1%, 2% and 3% in both directions with respect to the
reference value of the base-case, while all the other parameters were kept
unchanged.
Thus, 12 additional simulations, 6 for CALPUFF and 6 for SPRAY,
have been run, setting the source diameter as reported in Table 6.
These additional runs allowed for comparison of the sensitivity of the
SPRAY model (with 106′ 688′ 880 particles) and the CALPUFF model
related to variations of the source diameter.
Looking at the trend of the normalized sensitivity index on the plume
axis (Fig. 5), a first comment concerns the sign of the coefficient. For
CALPUFF, regardless of the considered receptor, the sensitivity coeffi­
cient obtained by the puff model is always negative, revealing that a
negative correlation between the input and the output exists. The
negative sign of the coefficient provides a clear indication that the
concentration value will be reduced because of an increment of the
diameter. The same general behaviour is detected in SPRAY, except for
few points, where a slightly positive value is obtained.
Another consideration concerns the general trend of the sensitivity
coefficient as a function of the source distance. Both models produce the
highest values (considering the absolute value) close to the source,
whereas values close to zero are identified for higher distances from the
source. This means that the investigated parameter has a greater influ­
ence on the receptors close to the emission source. This can be ascribed
to the plume rise mechanism, which is largely affected by the source
diameter: as previously discussed (Section 3.2), a change in the buoy­
ancy flux affects the pollutant dispersion close to the emission source
more than from far distances.
In addition, considering the receptors close to the source, higher
sensitivity coefficients are obtained when considering the point source
Table 6
Source diameter resulting from a variation of 1%, 2%, 3% of the average value.
Average value (base-case) = 5 m
− 1% +1% − 2% +2% − 3% +3%
4.95 m 5.05 m 4.90 m 5.10 m 4.85 m 5.15 m
11
Atmospheric Pollution Research 13 (2022) 101249
simulated by SPRAY than for the CALPUFF model. As discussed in
Section 3.1, the simulations with the CALPUFF - buoyant area source
model lead to a lower buoyancy flux compared to the point source
because of the heat losses. This in turn means that the plume rise phe­
nomenon is reduced, explaining why a variation in the source diameter
has less influence on the pollutant dispersion.
The obtained sensitivity indexes make it possible to identify whether
there is a linear relationship between the input and the output datum. As
reported by Yegnan et al. (2002), if the sensitivity coefficient is constant
over a range of input parameters (i.e. if there is not a change in the
sensitivity with a change in the input variable), the input datum can be
considered to be linear with respect to the output.
Concerning CALPUFF, as shown in Fig. 5, the 3 lines are so close
together that they almost overlap, indicating an almost linear depen­
dence of the ground level concentrations on the source diameter. Also,
the distance between the curves increases moving away from the source,
and this means that this linear dependence is gradually reduced.
On the other hand, in SPRAY simulations (run with 106′ 688′ 880
particles), the lines seem to be more distant from each other. However,
the scattered behaviour does not allow to clearly identify the distances
between the lines and, therefore, any comment concerning the input-
output relationship would be difficult.
Thus, to properly identify the relationship between the source
diameter and the pollutant concentration predicted by the two models,
some receptors, located at different distances from the source, have been
considered.
Considering, for instance, receptors placed at a distance of 1000 m
and 9000 m, respectively, from the source, the concentration trends
predicted by SPRAY and CALPUFF when changing the source diameter
are shown in Fig. 6:
In addition, in each plot, the trend line (linear type) is drawn, and the
resulting linear expression and the correlation coefficient (R2) are
displayed.
Considering the receptor located at 1000 m from the source, it can be
stated that the linear model properly approximates the input-output
relationship. This is true for both models, even if a higher R2 value is
found for CALPUFF.
By increasing the receptor distance from the source up to 9000 m, the
R2 coefficient decreases progressively for both models, even though a
better fit is still observed for the CALPUFF model. These considerations
are also confirmed by the analysis of other receptors, located at different
intermediate distances from the source (see Supplementary Material S5
for receptors located at 3000 m, 5000 m and 7000 m from the source).
The analysis on the individual receptors shows a decreasing reli­
ability of the linear relationship between output and input values when
moving away from the source. This may be justified considering the
wind effect on the dispersion phenomenon. Indeed, the Tracer concen­
trations modelled far from the source refer to the dispersion of the
pollutant that has been subjected to the wind field for longer distances.
The turbulent stochastic behaviour associated to the wind field pro­
motes the plume distortion, which will be more pronounced far from the
emission source. This observation is also confirmed, for instance, by the
ground level concentration maps resulting from the CALPUFF simula­
tions when changing the source diameter from 5 m (base-case) to 10 m.
The maps, reported in the Supplementary Material S6, show that, in the
vicinity of the source, the plume direction remains practically un­
changed, whereas at large distances, the wind effect results in a major
plume deviation.
Comparing the two models, it can be observed that, regardless of the
considered receptor, the R2 coefficients associated to the results of the
SPRAY model are always lower than those of the CALPUFF model. The
stochastic behaviour of the Lagrangian model influences the trajectory
of each particle: the component related to the turbulent fluctuation
provides to each particle a random character. This chaotic contribution
leads to a loss of linearity between the input and the output variables.
To conclude, among all the information provided by this analysis, the
F. Tagliaferri et al.
Fig. 5. Normalized sensitivity index as a function of the source distance resulting from SPRAY (106′ 688′ 880 particles) and CALPUFF simulations by changing the
source diameter by 1%, 2% and 3%.
Fig. 6. Maximum 1-h Tracer concentration as a function of the source diameter predicted by the CALPUFF model (left) and by the SPRAY model (right) on two
receptors located at 1000 m and 9000 m from the source, respectively.
most interesting is the one related to the different sensitivities of SPRAY
and CALPUFF to the investigated parameters. Indeed, the particle model
appears to have a higher numerical sensitivity with respect to the
diameter of the emission source. This discrepancy is particularly evident
in the vicinity of the emission source, where the sensitivity coefficients
resulting from SPRAY simulations are higher by one order of magnitude.
On the other hand, at larger distances the sensitivities of the two models
appear comparable.
4. Conclusions
When modelling the environmental effects of atmospheric pollution,
many sources of imprecision and uncertainty affect the results and
should therefore be critically analysed. Depending on the model
considered, there are numerous potential sources of variability, such as
the input data required by the model.
This paper aimed to compare the sensitivity of the SPRAY and the
CALPUFF models to input parameters when simulating the pollutant
dispersion from a hypothetical accidental fire, in order to identify the
most influential variables. In particular, the study focused on the effects
of input data regarding both source-term characterization and model-
specific parameters.
To this purpose, starting from a reference “base-case” scenario, other
alternative emission scenarios, characterized by a “macroscopic varia­
tion” of each variable, have been defined. For each alternative scenario,
three different source types have been studied: with CALPUFF the fire is
simulated as a buoyant area source, whereas with SPRAY the source is
12
Atmospheric Pollution Research 13 (2022) 101249
simulated as a point source and as a fire characterized by a 10% of
emitted particles having no buoyancy flux.
The most relevant outcome resulting from the investigation of the
alternative scenarios is that CALPUFF and SPRAY sensitivities to
“macroscopic variations” of the considered parameters are generally
comparable. The only significant difference is the model sensitivity to
the source diameter because, when using the specific option of the
SPRAY model to simulate emission from fires, the influence of the
diameter turns out to be negligible. Instead, the buoyant area source
modelled by CALPUFF and the point source simulated by SPRAY are
significantly affected by variations in the estimation of the source
diameter. Indeed, the simulations conducted at the boundaries of the
uncertainty range for the source diameter lead to variations of about
50–60% in the maximum Tracer concentrations at sensitive receptors.
On the other hand, when using SPRAY in combination with the specific
fire option, the source diameter variation gives a maximum variability of
10%. These discrepancies can be explained considering the different
relationships used to describe the plume rise phenomenon in the
different models (see Section 3.2). It should be highlighted that the
choice of the emission model that best approximates the real physical
behaviour is essentially left to the user.
All the other investigated variables concerning the source term
characteristics, such as source height and emission temperature, do not
significantly affect the model outputs, generally giving a maximum
variation in pollutant concentrations simulated at the receptors of about
10%. The model-specific parameters point out a non-controlling influ­
ence on the model results, generally producing variations of about 10%
F. Tagliaferri et al.
in the results as well.
Thus, in conclusion the most relevant parameters in terms of model
sensitivity turned out to be the source diameter and the amount of fuel
burnt, although the latter can usually be quantified with a certain degree
of reliability. Looking at the % variations relevant to the simulations of
the alternative scenarios, the source height seems the third most influ­
ential parameter, even though its influence on the model results is
limited (about 10%). The temperature of the emission source and the
investigated model-specific parameters lead to almost negligible
variations.
However, it is worth recalling that the diameter is the parameter that
leads to greater variability, but it is also one of those that has been varied
most in the alternative scenarios. This is because the estimation of the
source diameter is particularly critical and, consequently, a sufficiently
wide range of variability had to be considered.
Another observation arising from the first part is related to the dis­
crepancies obtained when using the different source options. In this
regard, modelling the fire as a point source is not recommended, since it
tends to underestimate ground concentrations. The use of the specific
fire options existing both for CALPUFF (buoyant area source) and SPRAY
(fire option), which generally produce comparable ground concentra­
tion trends as a function of the distance from the source (Fig. 2), is
actually recommended.
To effectively evaluate the model sensitivity to the source diameter
regardless of the reasonable range of variability, the sensitivity analysis
has been performed by applying a “microscopic variation” to this
parameter.
The most remarkable outcome resulting from this second approach is
that the SPRAY – point source model shows a significantly higher
sensitivity to the source diameter (of an order of magnitude) than
CALPUFF near the emission source. Conversely, at larger distances, the
sensitivity of the two models seems comparable.
This work allowed to evaluate the sensitivities of the SPRAY and
CALPUFF models to the investigated parameters from a theoretical and
numerical point of view. However, this paper is not intended to present a
sensitivity study applicable to any model or to any case study, but rather
it points out the importance of carrying out an investigation of the
possible range of variation of the input data in order to identify the most
influential variables. The obtained results can be useful to different
stakeholders (model users, environmental and control agencies) to have
a deeper knowledge of the possible range of variation of the simulated
ground concentration values deriving from the uncertainties in the
definition of the model input data.
It should be highlighted that the aim of this work is not to assess the
exactness of atmospheric dispersion models: to evaluate the accuracy of
the modelling results, in case of real accidental fires, the sensitivity
analysis should be coupled with some experimental validation in order
to evaluate the model capability to predict the experimental observa­
tions and, possibly, to improve and optimize its performances.
Declaration of competing interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.
Acknowledgements
The authors want to thank ARIANET for supporting the imple­
mentation of the case study with the SPRAY model.
Appendix A. Supplementary data
Supplementary data to this article can be found online at https://doi.
org/10.1016/j.apr.2021.101249.
13
Atmospheric Pollution Research 13 (2022) 101249
Credit author statement
Francesca Tagliaferri: Conceptualization, Methodology, Software,
Data curation, Writing – original draft, Visualization. Marzio Invernizzi:
Conceptualization, Data curation, Writing – review & editing, Visuali­
zation. Laura Capelli: Conceptualization, Visualization, Writing – review
& editing, Supervision.
Funding
This research did not receive any specific grant from funding
agencies in the public, commercial, or not-for-profit sectors.
References
Adame, J.A., Lope, L., Hidalgo, P.J., Sorribas, M., Gutiérrez-Álvarez, I., del Águila, A.,
Saiz-Lopez, A., Yela, M., 2018. Study of the exceptional meteorological conditions,
trace gases and particulate matter measured during the 2017 forest fire in Doñana
Natural Park, Spain. Sci. Total Environ. 645, 710–720. https://doi.org/10.1016/j.
scitotenv.2018.07.181.
Ainslie, B., Jackson, P.L., 2009. The use of an atmospheric dispersion model to determine
influence regions in the Prince George, B.C. airshed from the burning of open wood
waste piles. J. Environ. Manag. 90, 2393–2401. https://doi.org/10.1016/j.
jenvman.2008.11.009.
Antonioni, G., Burkhart, S., Burman, J., Dejoan, A., Fusco, A., Gaasbeek, R., Gjesdal, T.,
Jäppinen, A., Riikonen, K., Morra, P., Parmhed, O., Santiago, J.L., 2012. Comparison
of CFD and operational dispersion models in an urban-like environment. Atmos.
Environ. 47, 365–372. https://doi.org/10.1016/j.atmosenv.2011.10.053.
ARIANET, 2011. SURFPRO3 User’s Guide (SURFaceatmosphere Interface PROcessor,
Version 3). https://www.aria-net.it/it/.
Bhuiyan, A.A., Naser, J., 2015. Computational modelling of co-firing of biomass with
coal under oxy-fuel condition in a small scale furnace. Fuel 143, 455–466. https://
doi.org/10.1016/j.fuel.2014.11.089.
Björnham, O., Grahn, H., Burman, J., 2020. Comparison of the predictive results from the
two dispersion models PUMA and LPELLO with the JR II field data. Atmos. Environ.
233 https://doi.org/10.1016/j.atmosenv.2020.117521.
Booher, L.E., Janke, B., 1997. Air emissions from petroleum hydrocarbon fires during
controlled burning. Am. Ind. Hyg. Assoc. J. 58, 359–365. https://doi.org/10.1080/
15428119791012720.
BS EN 1363-2, 1999. Fire Resistance Tests. Part 2: Alternative and Additional
Procedures.
Cécé, R., Bernard, D., Brioude, J., Zahibo, N., 2016. Microscale anthropogenic pollution
modelling in a small tropical island during weak trade winds: Lagrangian particle
dispersion simulations using real nested LES meteorological fields. Atmos. Environ.
139, 98–112. https://doi.org/10.1016/j.atmosenv.2016.05.028.
Chang, J.I., Lin, C.C., 2006. A study of storage tank accidents. J. Loss Prev. Process. Ind.
19, 51–59. https://doi.org/10.1016/j.jlp.2005.05.015.
Chettouh, S., Hamzi, R., Innal, F., Haddad, D., 2014. Industrial fire simulation and
uncertainty associated with the Emission Dispersion Model. Clean Technol. Environ.
Policy 16, 1265–1273. https://doi.org/10.1007/s10098-014-0792-x.
Chutia, R., Mahanta, S., Datta, D., 2014. Uncertainty modelling of atmospheric
dispersion by stochastic response surface method under aleatory and epistemic
uncertainties. Sadhana - Acad. Proc. Eng. Sci. 39, 467–485. https://doi.org/
10.1007/s12046-013-0212-7.
Daly, A., Zannetti, P., Echekki, T., 2012. A combination of fire and dispersion modeling
techniques for simulating a warehouse fire. Int. J. Saf. Secur. Eng. 2, 368–380.
https://doi.org/10.2495/SAFE-V2-N4-368-380.
Devenish, B.J., Francis, P.N., Johnson, B.T., Sparks, R.S.J., Thomson, D.J., 2012.
Sensitivity analysis of dispersion modeling of volcanic ash from Eyjafjallajökull in
May 2010. J. Geophys. Res. Atmos. 117, 1–21. https://doi.org/10.1029/
2011JD016782.
DiNenno, P., Drysdale, D., Beyler, C., Walton, W., Custer, R., Hall, J., 2002. SFPE
Handbook of Fire Protection Engineering, third ed. (Quincy, Massachusetts).
Elbir, T., 2003. Comparison of model predictions with the data of an urban air quality
monitoring network in Izmir, Turkey. Atmos. Environ. 37, 2149–2157. https://doi.
org/10.1016/S1352-2310(03)00087-6.
Elbir, T., Mangir, N., Kara, M., Simsir, S., Eren, T., Ozdemir, S., 2010. Development of a
GIS-based decision support system for urban air quality management in the city of
Istanbul. Atmos. Environ. 44, 441–454. https://doi.org/10.1016/j.
atmosenv.2009.11.008.
Gant, S.E., Kelsey, A., McNally, K., Witlox, H., Bilio, M., 2013. Sensitivity analysis of
dispersion models for jet releases of dense-phase carbon dioxide. Chem. Eng. Trans.
31, 121–126. https://doi.org/10.3303/CET1331021.
Gonsamo, A., 2011. Normalized sensitivity measures for leaf area index estimation using
three-band spectral vegetation indices. Int. J. Rem. Sens. 32, 2069–2080. https://
doi.org/10.1080/01431161.2010.502153.
Griffiths, S.D., Chappell, P., Entwistle, J.A., Kelly, F.J., Deary, M.E., 2018. A study of
particulate emissions during 23 major industrial fires: implications for human
health. Environ. Int. 112, 310–323. https://doi.org/10.1016/j.envint.2017.12.018.
Hanna, S.R., Britter, R.E., 2002. Wind Flow and Vapor Cloud Dispersion at Industrial and
Urban Sites, Wind Flow and Vapor Cloud Dispersion at Industrial and Urban Sites.
https://doi.org/10.1002/9780470935613.
F. Tagliaferri et al.
Henderson, S.B., Burkholder, B., Jackson, P.L., Brauer, M., Ichoku, C., 2008. Use of
MODIS products to simplify and evaluate a forest fire plume dispersion model for
PM10 exposure assessment. Atmos. Environ. 42, 8524–8532. https://doi.org/
10.1016/j.atmosenv.2008.05.008.
Holmes, N.S., Morawska, L., 2006. A review of dispersion modelling and its application
to the dispersion of particles: an overview of different dispersion models available.
Atmos. Environ. 40, 5902–5928. https://doi.org/10.1016/j.atmosenv.2006.06.003.
Holnicki, P., Kałuszko, A., Trapp, W., 2016. An urban scale application and validation of
the CALPUFF model. Atmos. Pollut. Res. 7, 393–402. https://doi.org/10.1016/j.
apr.2015.10.016.
Holnicki, P., Nahorski, Z., 2015. Emission data uncertainty in urban air quality
modeling—case study. Environ. Model. Assess. 20, 583–597. https://doi.org/
10.1007/s10666-015-9445-7.
Ingason, H., Li, Y.Z., 2015. Tunnel Fire Dynamics. https://doi.org/10.1007/978-1-4939-
2199-7.
Invernizzi, M., Brancher, M., Sironi, S., Capelli, L., Piringer, M., Schauberger, G., 2020.
Odour impact assessment by considering short-term ambient concentrations: a multi-
model and two-site comparison. Environ. Int. 144, 105990. https://doi.org/
10.1016/j.envint.2020.105990.
Invernizzi, M., Tagliaferri, F., Sironi, S., Tinarelli, G., Capelli, L., 2021. Simulating
pollutant dispersion from accidental fires with a focus on source characterization.
J. Heal. Pollut. 11 https://doi.org/10.5696/2156-9614-11.30.210612.
Islam, M.A., 1999. Application of a Gaussian Plume model to determine the location of
an unknown emission source. Water. Air. Soil Pollut. 112, 241–245. https://doi.org/
10.1023/A:1005047321015.
Jang, C.B., Choi, S.W., Baek, J.B., 2015. CFD modeling and fire damage analysis of jet fire
on hydrogen pipeline in a pipe rack structure. Int. J. Hydrogen Energy 40,
15760–15772. https://doi.org/10.1016/j.ijhydene.2015.09.070.
Jung, Y.R., Park, W.G., Park, O.H., 2003. Pollution dispersion analysis using the puff
model with numerical flow field data. Mech. Res. Commun. 30, 277–286. https://
doi.org/10.1016/S0093-6413(03)00024-7.
Kota, S.H., Ying, Q., Zhang, Y., 2013. Simulating near-road reactive dispersion of gaseous
air pollutants using a three-dimensional Eulerian model. Sci. Total Environ. 454
(455), 348–357. https://doi.org/10.1016/j.scitotenv.2013.03.039.
Langmann, B., Duncan, B., Textor, C., Trentmann, J., van der Werf, G.R., 2009.
Vegetation fire emissions and their impact on air pollution and climate. Atmos.
Environ. 43, 107–116. https://doi.org/10.1016/j.atmosenv.2008.09.047.
Leelőssy, Á., Molnár, F., Izsák, F., Havasi, Á., Lagzi, I., Mészáros, R., 2014. Dispersion
modeling of air pollutants in the atmosphere: a review. Cent. Eur. J. Geosci. 6,
257–278. https://doi.org/10.2478/s13533-012-0188-6.
Lemieux, P.M., Lutes, C.C., Santoianni, D.A., 2004. Emissions of organic air toxics from
open burning: a comprehensive review. Progress in Energy and Combustion Science.
https://doi.org/10.1016/j.pecs.2003.08.001.
Liu, T., Marlier, M.E., DeFries, R.S., Westervelt, D.M., Xia, K.R., Fiore, A.M., Mickley, L.
J., Cusworth, D.H., Milly, G., 2018. Seasonal impact of regional outdoor biomass
burning on air pollution in three Indian cities: Delhi, Bengaluru, and Pune. Atmos.
Environ. 172, 83–92. https://doi.org/10.1016/j.atmosenv.2017.10.024.
Liu, X., Godbole, A., Lu, C., Michal, G., Venton, P., 2015. Optimisation of dispersion
parameters of Gaussian plume model for CO2 dispersion. Environ. Sci. Pollut. Res.
22, 18288. https://doi.org/10.1007/s11356-015-5404-8, 18299.
Markatos, N.C., Christolis, C., Argyropoulos, C., 2009. Mathematical modeling of toxic
pollutants dispersion from large tank fires and assessment of acute effects for fire
fighters. Int. J. Heat Mass Tran. 52, 4021–4030. https://doi.org/10.1016/j.
ijheatmasstransfer.2009.03.039.
Mishra, K.B., Wehrstedt, K.D., 2015. Underground gas pipeline explosion and fire: CFD
based assessment of foreseeability. J. Nat. Gas Sci. Eng. 24, 526–542. https://doi.
org/10.1016/j.jngse.2015.04.010.
Nivolianitou, Z., Konstandinidou, M., Michalis, C., 2006. Statistical analysis of major
accidents in petrochemical industry notified to the major accident reporting system
(MARS). J. Hazard Mater. 137, 1–7. https://doi.org/10.1016/j.
jhazmat.2004.12.042.
Novozhilov, V., 2001. Computational fluid dynamics modeling of compartment fires.
Prog. Energy Combust. Sci. 27, 611–666. https://doi.org/10.1016/S0360-1285(01)
00005-3.
14
Atmospheric Pollution Research 13 (2022) 101249
Ravina, M., Panepinto, D., Zanetti, M.C., 2020. Development of the DIDEM Model:
comparative evaluation of CALPUFF and SPRAY dispersion models. Int. J. Environ.
Impacts Manag. Mitig. Recover. 3, 1–18. https://doi.org/10.2495/ei-v3-n1-1-18.
Rodrigues, H.S., Monteiro, M.T.T., Torres, D.F.M., 2013. Sensitivity analysis in a dengue
epidemiological model. Conf. Pap. Math. 1–7. https://doi.org/10.1155/2013/
721406, 2013.
Russell, A., Dennis, R., 2000. NARSTO critical review of photochemical models and
modeling. Atmos. Environ. 34, 2283–2324. https://doi.org/10.1016/S1352-2310
(99)00468-9.
Rzeszutek, M., 2019. Parameterization and evaluation of the CALMET/CALPUFF model
system in near-field and complex terrain - terrain data, grid resolution and terrain
adjustment method. Sci. Total Environ. 689, 31–46. https://doi.org/10.1016/j.
scitotenv.2019.06.379.
Santiago, J.L., Martín, F., 2008. SLP-2D: a new Lagrangian particle model to simulate
pollutant dispersion in street canyons. Atmos. Environ. 42, 3927–3936. https://doi.
org/10.1016/j.atmosenv.2007.05.038.
Scire, J.S., Strimaitis, D.G., Yamartino, R.J., 2000. A user’s guide for the CALPUFF
dispersion model. Earth Tech. Inc 521.
Seibert, P., 2000. Uncertainties in atmospheric dispersion modelling and source
determination. In: Procedings Informal Work. Meteorol. Model. Support CTBT,
Vienna, pp. 1–4.
Seland, Oø, Iversen, T., 1999. A scheme for black carbon and sulphate aerosols tested in a
hemispheric scale, Eulerian dispersion model. Atmos. Environ. 33, 2853–2879.
https://doi.org/10.1016/S1352-2310(98)00389-6.
Shie, R.H., Chan, C.C., 2013. Tracking hazardous air pollutants from a refinery fire by
applying on-line and off-line air monitoring and back trajectory modeling. J. Hazard
Mater. 261, 72–82. https://doi.org/10.1016/j.jhazmat.2013.07.017.
Sinha, N., Ghose, M.K., Singh, G., Srivastava, S., Sinha, I.N., 2004. Classification of air
pollution dispersion models : a critical review. Proc. Natl. Semin. Environ. Eng. with
Spec. Emphas. Min. Environ. 2004–19.
Sonnemans, P.J.M., Körvers, P.M.W., Pasman, H.J., 2010. Accidents in “normal”
operation - can you see them coming? J. Loss Prev. Process. Ind. 23, 351–366.
https://doi.org/10.1016/j.jlp.2010.01.001.
Souto, M.J., Souto, J.A., Pérez-Muñuzuri, V., Casares, J.J., Bermúdez, J.L., 2001.
A comparison of operational Lagrangian particle and adaptive puff models for plume
dispersion forecasting. Atmos. Environ. 35, 2349–2360. https://doi.org/10.1016/
S1352-2310(00)00537-9.
Srinivas, C.V., Hari Prasad, K.B.R.R., Naidu, C.V., Baskaran, R., Venkatraman, B., 2016.
Sensitivity analysis of atmospheric dispersion simulations by FLEXPART to the WRF-
simulated meteorological predictions in a coastal environment. Pure Appl. Geophys.
173, 675–700. https://doi.org/10.1007/s00024-015-1104-z.
Tinarelli, G., 2017. SPRAY 3 . 1 - General Description and User ’ S Guide.
UNI 10796, 2009. Valutazione della dispersione in atmosfera di effluenti aeriformi. http:
//store.uni.com/catalogo/uni-10796-2000.
UNI 10964, 2009. Guida alla selezione dei modelli matematici per la previsione di
impatto sulla qualità dell ’ aria evaluation of air quality impact. http://store.uni.co
m/catalogo/uni-10964-2001?josso_back_to=http://store.uni.com/josso-security-ch
eck.php&josso_cmd=login_o.
US EPA, 1998. AP-42, “Compilation of Air Emission Factors” Chapter 1: External
Combustion Sources. https://www.epa.gov/nscep.
US EPA, 1995. B-95, User’s Guide for the Industrial Source Complex (ISC3) Dispersion
Models - Volume Ii - Description of Model Algorithms, fifth ed. https://www.epa.
gov/nscep.
Weichenthal, S., Van Rijswijk, D., Kulka, R., You, H., Van Ryswyk, K., Willey, J.,
Dugandzic, R., Sutcliffe, R., Moulton, J., Baike, M., White, L., Charland, J.P.,
Jessiman, B., 2015. The impact of a landfill fire on ambient air quality in the north: a
case study in Iqaluit, Canada. Environ. Res. 142, 46–50. https://doi.org/10.1016/j.
envres.2015.06.018.
Yegnan, A., Williamson, D.G., Graettinger, A.J., 2002. Uncertainty analysis in air
dispersion modeling. Environ. Model. Software 17, 639–649. https://doi.org/
10.1016/S1364-8152(02)00026-9.
Zheng, B., Chen, G., 2011. Storage tank fire accidents. Process Saf. Prog. 30, 291–293.